Improving Generalization for AI-Synthesized Voice Detection

TL;DR

The paper tackles the poor cross-domain generalization of AI-synthesized voice detectors by identifying that reliance on vocoder-specific artifacts and sharp loss landscapes limit robustness. It introduces a disentanglement framework that separates domain-specific and domain-agnostic artifacts and aligns domain-agnostic features with content via mutual information, complemented by reconstruction regularization. Optimization employs sharpness-aware minimization to flatten the loss landscape and promote stable generalization. Across LibriSeVoc, WaveFake, ASVspoof2019, and FakeAVCeleb, the approach achieves state-of-the-art generalization, with notable gains in unseen-vocoder scenarios, highlighting its potential for real-world fake-audio detection.

Abstract

AI-synthesized voice technology has the potential to create realistic human voices for beneficial applications, but it can also be misused for malicious purposes. While existing AI-synthesized voice detection models excel in intra-domain evaluation, they face challenges in generalizing across different domains, potentially becoming obsolete as new voice generators emerge. Current solutions use diverse data and advanced machine learning techniques (e.g., domain-invariant representation, self-supervised learning), but are limited by predefined vocoders and sensitivity to factors like background noise and speaker identity. In this work, we introduce an innovative disentanglement framework aimed at extracting domain-agnostic artifact features related to vocoders. Utilizing these features, we enhance model learning in a flat loss landscape, enabling escape from suboptimal solutions and improving generalization. Extensive experiments on benchmarks show our approach outperforms state-of-the-art methods, achieving up to 5.12% improvement in the equal error rate metric in intra-domain and 7.59% in cross-domain evaluations.

Figures (5)

• Figure 1: Comparison of audio deepfake detection methods. (a) The conventional method uses entire voice features and excels in distinguishing between human and AI-synthesized voices within familiar domains but struggles with voices generated from unseen vocoders, leading to inaccurate separation. (b) Our method achieves intra-&cross-domain detection by exposing domain-agnostic features for learning on a flattened loss landscape.
• Figure 2: Experimental results for Motivation. (Left) The differences of mel-spectrogram between human voices and AI-synthesized ones (e.g., based on WaveGrad chen2020wavegrad and WaveRNN kalchbrenner2018efficient vocoders). The red circles highlight AI vocoder artifacts. More details about these differences can be found in Appendix. (Middle) The UMAP 2018arXivUMAP visualization of features from related methods and our framework on LibriSeVoc sun2023ai. The genuine and forged voices from six vocoders are separated in the latent space. The baselines (RawNet2 tak2021end, Sun et al.sun2023ai) and our domain-specific module can learn the domain-specific features, whereas the domain-agnostic module of our method captures the shared forgery features across different vocoders, and the content module exclusively captures forgery-irrelevant features. (Right) Visualization of loss landscape for RawNet2 and Sun et al. The sharp local and global minima could lead to models with poor generalization.
• Figure 3: The overall architecture of our proposed approach. (a) In the encoder module, two RawNet2 tak2021end-structured backbones are added to the audio data signal to extract content and artifact features. Additionally, two headers further categorize artifact features into domain-agnostic and domain-specific categories. (b) In the decoder module, we use the content features as the base and fuse them with their own forgery features, as well as those from other samples for audio reconstruction. (c) Domain-agnostic features are made universally applicable by maximizing mutual information between domain-agnostic features and content features through joint and marginal distributions. (d) The sharpness-aware minimization (SAM) serves as an optimization technique to guide the model toward a flatter loss landscape to enhance its generalization. (e) For the inference, we take the predicted results of the domain-agnostic classification header.
• Figure 4: (Left-Top) The virtualizations of model with MI ($V_c$) and model without MI ($V_d$). (Left-Bottom) The loss landscape visualization of our method with and without flattening the loss landscape. (Middle) (a) The effect of $\lambda_4$ for balancing mutual information term. (b) The effect of $\gamma$ in sharpness-aware minimization. (Right) The effect of the number of vocoders used in the train set. The reported average Equal Error Rate for seen (aEERs) and average Equal Error Rate for unseen (aEERu) domains.
• Figure 5: The artifacts introduced by the AI vocoders to a human voice signal. We show the mel-spectrogram of these signals and their mutual differences by making the left mel-spectrogram minus the right ones, respectively